$1M DARPA Grant for Signal Processing in Neural Networks for Wireless IoT
Abstract: ECE Associate Professor Kaushik Chowdhury, Assistant Professor Pau Closas, Professor Deniz Erdogmus, Professor Tommaso Melodia, and Assistant Professor Yanzhi Wang received $1M funding from DARPA for their project titled Signal Processing in Neural Networks (SPiNN) for Wireless IoT. The team will create deep neural network kernels that will replace and outperform models and iterative algorithms for four typical components in a wireless communication receiver: channel equalization, denoising, demodulation, and error correction decoding. To this end, variational autoencoders and generative adversarial networks will be used to produce realistic data to complement real data. The trained neural network kernels will be optimally compressed for FPGA implementation to meet real-time computing requirements in hardware.
Full Story: A powerhouse team of machine learning, signal processing and wireless communications experts led by Electrical and Computer Engineering Professor Deniz Erdogmus is collaborating on a project that aims to significantly improve wireless communications in our increasingly interconnected world.
Erdogmus and his fellow ECE researchers, Associate Professor Kaushik Chowdhury, Assistant Professor Pau Closas, Professor Tommaso Melodia and Assistant Professor Yanzhi Wang, will partner on an 18-month, $1 million contract recently awarded by the Defense Advanced Research Projects Agency (DARPA) for Signal Processing in Neural Networks (SPINN) for Wireless Internet of Things (IOT).
“DARPA wants a solution at a high level for wireless communications and radar systems using deep neural networks (DNNs) or deep learning-based solutions,” says Erdogmus. “We chose to go for the wireless communications application because it’s an area of strength for our team.”
Neural networks are computational mathematical models which mirror the organization of the human brain and its millions of processing units. “We can solve really complex problems by appropriately training these neural networks,” explains Erdogmus. “For wireless communication receivers, we will take classical module units and replace them with DNNs in Phase 1 of the project and then improve those neural network solutions in Phase 2.” The team will use DNN models to enhance three basic receiver components: channel estimation and equalization; demodulation; and error correction decoding.
Pioneering a new approach
Erdogmus notes that while the concept of using DNNs in wireless communications is not new, the team’s approach is novel. “What is unique is the way we’re actually constructing the DNN architecture we’ll be using and the way we’re training the networks,” he says.
According to Erdogmus, there is typically not enough data to train large neural network models for practical applications. The team’s proposed solution is to build data systems to produce realistic data that imitate real data. “We’ll take real data from Northeastern’s Institute for the Wireless Internet of Things, set up an infrastructure, then train a generative transmitter/channel model of this Wi-Fi setup so we can produce any amount of realistic data,” he says. “Basically, the innovations we’re bringing here on the receiver side are the design of neural networks and choices about what type of structure; on the training side, we will be looking at realistic data synthesizers that can produce real data, so that we can cut down on the need for real data acquisition, which is an expensive, cumbersome process.”
Ultimately the benefit of the team’s research will be better wireless communications through higher data transmission rates with lower power consumption. “That means your cellphone can stream Netflix more effectively,” Erdogmus explains. “That’s basically what we’re trying to accomplish. You need higher communications bandwidth to support more users and be able to transmit more data.”
Erdogmus sees the DARPA project as an important step in building on the strength of Northeastern’s machine learning expertise and its Institute for the Wireless Internet of Things, a leader in advancing the wireless systems and networks of the future. “I expect this project will lead to other collaborations between machine learning and wireless communications researchers.”